Automation,OccupationalEarningsTrends ...zachparolin.com/wp-content/uploads/2020/05/SocialForces_Parolin.pdfreviewers.I also appreciate the feedback of participants at the 2018 ESPANet

Automation, Occupational Earnings Trends, and the Moderating Role of Organized Labor 1

Automation, Occupational Earnings Trends, and the Moderating Role of Organized Labor

Automation, Occupational Earnings Trends, and theModerating Role of Organized Labor

Zachary Parolin, Columbia University

Routine-biased technological change has emerged as the dominant explanationfor the differential earnings growth of occupations at greater risk of automation,such as machine operators or office clerks, relative to less routine occupations.

In contrast, this paper finds that the declining earnings returns to an occupation’sroutine task intensity (RTI) can largely be attributed to the decline of organized labor.Using individual-level data on 3.3 million employed adults across the United States from1983 to 2017, this paper finds that organized labor has two countervailing effects onoccupations at greater risk of automation. First, higher union coverage within a stateand industry inhibits the decline in earnings returns to an occupation’s RTI. Second,higher union coverage hastens the decline in employment shares of occupations withhigher RTI. The result is that occupations at greater risk of automation experience morefavorable earnings growth where unions are more resilient, but at the cost of acceler-ated declines in their employment shares. Counterfactual analyses demonstrate that ifunion coverage in the United States had remained stable at 1983 levels, the earningsreturns to an occupation’s RTI might not have declined from 1983 to 2017, and theobserved pattern of occupational earnings polarization in the 1990s might not haveoccurred. However, the mean RTI of occupations might have declined by an additional21 percent from 1983 to 2017 relative to the observed decline. The findings suggest thatthe social consequences of automation are conditional on the strength of organizedlabor.

Introduction

The effect of technological progress on the structure of employment and earningshas emerged as a central concern of labor market research. A dominant per-spective in recent literature holds that advancements in technology contribute

This paper was completed as part of an OECD Future of Work Fellowship. For helpful commentsand suggestions, I am grateful to Andrea Salvatori, Stijn Broecke, Diego Collado, Andrew Green,Dario Guarascio, Wim Van Lancker, Brian Nolan, Jose Pacas, Linus Sioland, and four anonymousreviewers. I also appreciate the feedback of participants at the 2018 ESPANet Conference in Vilnius,the “What Drives Inequality?” workshop in Luxembourg, the Herman Deleeck Centre for SocialPolicy Lunch Seminar series, the “Labor Market Liberalization after the Lehman Crisis” workshopin Tokyo, and the 2019 SASE Annual Meeting; e-mail: .

© The Author(s) 2020. Published by Oxford University Press on behalf of the Universityof North Carolina at Chapel Hill. All rights reserved. For permissions, please e-mail:[email protected].

Social Forces 00(00) 1–26doi:10.1093/sf/soaa032

Dow

nloaded from https://academ

ic.oup.com/sf/advance-article-abstract/doi/10.1093/sf/soaa032/5835499 by zachparolin@

gmail.com

on 14 May 2020

2 Social Forces

to differential earnings and employment growth in routine occupations, suchas machine operators or office clerks, relative to less routine occupations, suchas managers or service workers. Indeed, evidence suggests that declining demandfor routine tasks has contributed to declining employment shares of automatableoccupations across a broad range of advanced economies (Goos, Manning, andSalomons 2014). With respect to earnings growth, however the story is mixed.During the 1990s, the earnings of automatable occupations in the United Statesgrew at a slower rate than less routine occupations at the bottom and top of thewage distribution (Autor and Dorn 2013; Mishel, Schmitt, and Shierholz 2013).Outside of the United States, however, this pattern of occupational earningspolarization has generally not been observed (Naticchioni, Ragusa, and Massari2014). In fact, earnings growth of automatable occupations continues to outpacethat of low-routine, low-pay occupations in many advanced economies.

One defining difference of European and American labor markets is thestrength of organized labor. While prior work has demonstrated that organizedlabor plays a central role in compressing the earnings distribution and increasinglabor’s share of national income (e.g., Brady, Baker, and Finnigan (2013),Kristal (2013), Western and Rosenfeld (2011)), it remains unclear the extentto which declining worker power can explain patterns of occupational earningspolarization. This study investigates how labor unions moderate the effect oftechnological change on the relative earnings growth of occupations with higherroutine task intensity (RTI). To do so, I apply Autor and Dorn’s (2013) taskcomposition data to a cross-state US sample of 3.3 million employed adultsusing the US Current Population Survey Merged Outgoing Rotation Groups(CPS MORG).

In contrast to the structural and macroeconomic explanations of occupationalearnings trends, which ascribe changes in relative earnings growth of automat-able occupations to exogenous shifts in the demand for routine tasks (Acemogluand Autor 2011; Autor, Levy, and Murnane 2003; Autor and Dorn 2013), Iposit that earnings and employment trends among occupations at greater riskof automation are conditional on the strength of organized labor. Specifically, Iinvestigate the existence of an earnings effect, in which higher union membershipwithin a state-industry inhibits declining earnings returns to an occupation’s RTI.I also investigate the possibility of an employment share effect, in which unionmembers can protect the earnings growth of high RTI occupations, but at thecost of an accelerated decline in employment shares among such occupations.The existence of these two countervailing effects would suggest that the strengthof organized labor shapes patterns of occupational earnings and employmentpolarization, two phenomena that have been central to emergent literature onautomation and technological change.

Findings from state-year-industry fixed effects models support the earningsand employment share effect hypotheses. A counterfactual simulation demon-strates that had union coverage in the United States remained stable from1983 to 2017, the earnings returns to an occupation’s RTI might not havedeclined between 1983 and 2017, rather than experiencing a 0.30 log pointdecline. Moreover, the observed pattern of occupational wage polarization in

Dow



gmail.com

on 14 May 2020


the United States during the 1990s might not have occurred. However, the higherrelative earnings growth for high RTI occupations appears to come at the costof accelerated declines in employment shares. Had union coverage remainedconstant, the mean RTI of occupations might have fallen by an additional 21percent relative to the observed decline from 1983 to 2017. The findings suggestthat theories connecting automation to occupational earnings trends must berooted in political-institutional context. Organized labor appears to shape thesocial consequences of technological change.

Background and Theory

Automation and Occupational Earnings

Technological change has been central to analyses of the earnings distributionthroughout recent decades. In the 1990s, labor market scholars advanced theconcept of skill-biased technological change (SBTC), the idea that technologicaladvancements strengthen the wage premium for workers with a college degreedespite the rising supply of such workers (Katz and Murphy 1992). As aptlysummarized in the title of Goldin and Katz’s (2008) book, The Race BetweenEducation and Technology, the SBTC framework views rising inequality throughthe lens of technology-propelled demand for more educated workers. As severalsociologists have documented, the SBTC framework overlooks the importance ofworker power in its diagnosis of rising inequality (e.g., Fernandez (2001), Kristal(2013)), but the framework also suffers from another flaw: it cannot explainwhy low-pay, service-sector occupations were experiencing rising earnings andemployment shares relative to middle-pay occupations throughout the 1990s.These patterns of employment and wage polarization led a separate groupof scholars to abandon the “skills” framework and instead adopt a “tasks-based” framework, focusing on the routineness of tasks performed by a givenoccupation (Acemoglu and Autor 2011; Autor and Dorn 2013; Goos, Manning,and Salomons 2009). In contrast to SBTC, routine-biased technological change(RBTC) could explain “the heterogeneous behavior of the top, middle andbottom of the earnings distribution” (Mishel et al. 2013).

Specifically, the RBTC framework suggests that technological progress gen-erates two primary consequences with respect to employment and earningspatterns (Acemoglu and Autor 2011; Autor et al. 2003; Goos et al. 2014).One is the polarization of employment structures: high RTI occupations, themost susceptible to automation, tend to be in the middle of the earnings andskill distributions (Acemoglu and Autor 2011; Dwyer 2013). Their decline thusleads to a dip in employment shares near the median of the earnings distri-bution. Meanwhile, technological advancements are theorized to take a factor-augmenting form for higher-skill occupations, contributing to rising employmentshares at the top the earnings and skill distributions (Autor et al. 2003; Autor andDorn 2013).1 Though the decline in employment shares of high RTI occupationshas been written about extensively in the context of the United States, evidence

Dow



gmail.com

on 14 May 2020

4 Social Forces

suggests that the trend is also pervasive throughout other advanced economies(Goos et al. 2009; OECD 2017).

A second potential consequence of RBTC is that earnings of high RTI occu-pations may experience less favorable growth relative to low RTI occupations(Acemoglu and Autor 2011; Goos et al. 2009). If declining demand for routinetasks translates into declining relative earnings among high RTI occupations,then earnings polarization, in addition to employment polarization, might occur(Böhm 2017; Firpo, Fortin, and Lemieux 2011). In the United States during the1990s, for example, the earnings of occupations at the bottom and top of thewage distribution increased at a greater rate than earnings of occupations in themiddle of the distribution (Acemoglu and Autor 2011). However, evidence fromother advanced economies (and other decades within the United States) suggeststhat earnings polarization is far from the norm. In Germany, for example, occu-pational polarization has occurred without earnings polarization (Antonczyk,DeLeire, and Fitzenberger 2010). Looking more broadly at a number of EUMember States, Naticchioni et al. (2014) also find that technological change hasnot led to the polarization of earnings. In other words, job polarization appearsto be pervasive across advanced economies, but earnings polarization has oftennot followed.

The inconsistency in the relative earnings growth of automatable occupationsposes prima facie evidence against an important feature of the RBTC hypothesis.As Acemoglu and Autor (2011) describe, a central innovation of their “taskframework” is its ability to simultaneously explain the decline of earnings in highRTI occupations and the rise of earnings in low-pay and high-pay occupations(wage polarization) (Mishel et al. 2013). That earnings trends across manyadvanced economies do not align with this pattern raises an important question:under which conditions do the earnings of high RTI occupations continue togrow at a faster rate than less routine occupations?

One shortcoming of the RBTC hypothesis, as well as other estimations ofthe social consequences of automation, is a failure to appropriately accountfor the role of labor market institutions, and the strength of organized laborin particular, in shaping the effects of automation on occupational earningstrends. Put differently, this paper argues that the dominant practice of rootingtheories of occupational earnings growth solely in structural terms overlooks themoderating role of organized labor.

Labor Market Institutions and Occupational Earnings Trends

Labor market institutions have not been at the forefront of the RBTC literature.This is perhaps due to an extensive focus on a single country (the United States)or perhaps due to a greater focus on the employment effects of technologicalchange. Regardless of the cause, the relegation of institutions, and the role oforganized labor in particular, is surprising given the rich history of literature ontheir role in shaping the earnings distribution within and between occupations(Biegert 2017; Brady 2009; Katz and Autor 1999; OECD 2011; VanHeuvelen2018; Western and Rosenfeld 2011).

Dow



gmail.com

on 14 May 2020


How might organized labor affect the social consequences of technologicalchange? Building on prior sociological research relating to power resource theoryand comparative institutions, I hypothesize that higher union membership withina state-industry inhibits declining earnings returns to an occupation’s RTI. Irefer to this as an earnings effect. Building on literature related to potentialemployment share effects of unionization, I then propose that organized laborcan achieve higher earnings gains for high RTI occupations, but at the cost ofan accelerated decline in employment shares. I describe the evidence supportingthese two hypotheses in turn.

With respect to the earnings growth of high RTI occupations, it is firstworth reiterating the baseline perspective in the RBTC literature. In a perfectlycompetitive market, economic theory suggests that the declining demand forroutine tasks will generally translate into lower earnings for occupations atgreater risk of automation, lest the individuals in the occupations lose their jobsaltogether (Acemoglu and Autor 2011). However, the comparative literature onlabor market institutions suggests that several contextual factors might impede astraightforward relationship between declining demand and declining earnings.Labor unions and collective bargaining, in particular, play a critical role inshaping the distribution of market earnings. From a cross-national perspective,high levels of bargaining coverage have been linked to more compressed earningsdistributions (Hirsch 2004; OECD 2017; Visser and Checchi 2011). From across-state perspective within the United States, higher levels of unionizationhave been linked to lower levels of in-work poverty (Brady et al. 2013), and moreegalitarians pay norms (VanHeuvelen 2018; Western and Rosenfeld 2011).

Power resource theory, central to the comparative institution literature, alsopoints to the role of organized labor as central actor in a broader, class-basedstruggle over the distribution of resources (Brady 2009; Brady, Blome, andKleider 2016; Jacobs and Dirlam 2016; Korpi 1983; Korpi 1985; Wilmers 2017).Most directly, unions and collective bargaining agreements generally contributeto higher earnings for covered workers, as well as improvements in workingconditions (Brady et al. 2016; Kalleberg, Wallace, and Raffalovich 1984; Kristal2013). Unions have lifted earnings across a broad range of occupation types, butthe earnings effects have historically been stronger among workers in industrialoccupations (DiNardo, Fortin, and Lemieux 1996). Such occupations tend to bein the middle of the earnings distribution and tend to have a higher RTI; as such,they have been the particular focus of the earnings polarization literature (Autorand Dorn 2013). In addition to contributing to earnings increases for unionizedworkers, unions also have spillover effects for non-unionized workers (Deniceand Rosenfeld 2018). Thus, the effects of unionization on a worker’s earningsgrowth cannot be understood solely as a function of the individual’s union status,but should also take into account the union membership of other workers in theperson’s firm or industry.

The bargaining power that a trade union provides is likely to be particularlyimportant when a given industry or occupation is under threat. Where organizedlabor is weakest, market forces are more likely to exert downward pressureon the earnings of high RTI occupations (Nickell and Andrews 1983). Where

Dow



gmail.com

on 14 May 2020

6 Social Forces

organized labor has greater bargaining power, such occupations are perhapslikely to experience more favorable earnings growth. This possible earningseffect suggests that higher union membership within a state-industry inhibitsdeclining earnings returns to an occupation’s RTI. Put differently, declining unionmembership contributes to declining relative earnings growth for occupations atgreater risk of automation.

Might the enhanced earnings growth for union members, however, come at acost? This study’s second hypothesis posits that higher union membership withina state-industry is also associated with accelerated declines in the employmentshares of high-routine occupations. Put differently, this potential employmentshare effect suggests that the average RTI of occupations within a state-industrydeclines at a faster rate when union membership is higher. This hypothesis isrooted in related bodies of research that have worked to understand the potentialemployment effects of unionization (Leonard 1992; Lindbeck and Snower 2001;Nickell and Andrews 1983; Pencavel 1984) and the role of labor costs in shapingthe decline of routine occupations (Lordan and Neumark 2017).

At the core of a potential employment share effect is evidence that in a contextof declining demand for routine tasks, the rate at which routine occupationsexperience declining employment shares is conditional on their relative laborcosts. As one example, Lordan and Neumark (2017) find that higher labor costsdue to minimum wage increases reduce the share of high-routine jobs in a givenstate and industry. Moreover, rising labor costs contributed to “increases in thelikelihood that low-skilled workers in automatable jobs become nonemployedor employed in worse jobs.” Consider also the relative growth of high-routinejobs in states with lower labor costs and anti-union right-to-work (RTW) laws inrecent decades (Newman 1983). Foreign auto manufacturers, such as Hyundai,Nissan, and Toyota, have opened US plants and created routine jobs almostexclusively in states across the American South—a region where wages are lowerand where organized labor has struggled to maintain a presence (Rosenfeld2014).2 Internationally, the tension among declining demands, earnings, andemployments was also present in German manufacturing industries during therecent financial crisis. Recognizing that declining consumption might otherwiseforce large layoffs, many German workers participated in the “short-time work”program, reducing their hours (with subsidies from the state) to reduce theirfirm’s aggregate labor costs. Without the temporary reduction in labor costs,“unemployment would have risen by approximately twice as much as it actuallydid” (Brenke, Rinne, and Zimmermann 2011).

Given this evidence of trade-offs between employment and earnings forroutine occupations, and the fact that routine occupations already face a risingemployment penalty (Goos et al. 2009), I expect that higher shares of unionmembership, insofar as they are successful in protecting the relative earningsof routine occupations, will amplify the rising employment penalty for suchoccupations. Importantly, this does not necessarily imply that unionized, routineworkers bear the costs of the accelerated decline in routine employment shares.Instead, declines in the hiring of routine occupations (relative to lower RTIoccupations) could plausibly drive the aggregate decline in the average RTI of

Dow



gmail.com

on 14 May 2020


occupations within an industry. This pattern would be consistent with resultsfrom Cortes (2016) who, in a panel study of the employment trajectories ofroutine workers, finds that the decline of routine jobs in the US labor marketis largely due to reduced hiring of routine workers rather than accelerationsin layoffs. Similarly, research from Fernandez (2001), Kimeldorf (1992), andAhlquist and Levi (2013) find that unions negotiate the introduction of labor-replacing technologies in a way that dampens their own employment losses.Thus, higher union membership might enhance the employment protectionsfor unionized members, but contribute to (1) declines in hiring of high RTIoccupations within the state-industry relative to the pace of hiring in state-industries with lower labor costs (consistent with Lordan and Neumark (2017))and (2) declines in the hiring of high RTI occupations within the state-industryrelative to the pace of hiring for low RTI occupations (consistent with Cortes(2016)). The net effect of either of these outcomes is a faster decline in the averageRTI of occupations where union membership is more resilient.

The earnings and employment effects may be interconnected. With fewerjob opportunities available, a jobseeker who would otherwise pursue a highRTI occupation may decide to pursue a lower-pay occupation instead, furtherdecreasing the share of high RTI occupations relative to the full distributionof jobs. Increased competition for these lower-pay jobs might inhibit earningsgrowth for such occupations and help to ensure that the relative earnings ofroutine jobs continue to grow at a faster pace. Thus, an employment share effectmight reinforce the hypothesized earnings effect. Indeed, Autor and Dorn (2013)speculate that the decline of wages in low-pay, low RTI occupations in the UnitedStates in recent years is partly due to more “middle-skill” workers competing forthe low-pay occupations.

Potential Objections

There are at least three explanations for why empirical evidence may notsupport this study’s hypotheses. First, variation in trends of state-industry unionmembership may be too small to reveal meaningful effects on the relativeearnings trends. Consider that levels of unionization have declined across allstates in the United States in recent decades (Hirsch and MacPherson 2003).However, descriptive statistics from the Current Population Survey (discussed inmore detail in the next section) suggest large variation in rates of union declineacross the 50 states. In Vermont, for example, union membership declined from17.5 percent in 1983 to 15.3 percent in 2015, a 13 percent decline. In Utah, bycontrast, union membership fell from 23.1 percent to 6.3 percent, a 73 percentdecline, over the same timeframe. The average, unweighted decline across the 50states was 44 percent with a standard deviation (SD) of 12.3 percent.

Second, there may be some relationship between technological change anddecline in union membership, which could affect this study’s focus on therelationship between union membership and the earnings returns to an occu-pation’s RTI. Later, I detail how I account for this possibility in my estimationstrategy. However, it is worth noting that prior research suggests that political

Dow



gmail.com

on 14 May 2020

8 Social Forces

and institutional forces can be as important as structural forces in explainingthe decline of unionization. Anti-union laws, such as RTW legislation, andregulations restricting union formation are key contributors of organized labor’songoing decline and perhaps explain more of the decline in union membershipthan any recent acceleration in labor-replacing technologies (Hertel-Fernandez2019; Thelen 2015). Descriptive evidence from the CPS again supports theseclaims. In RTW states, union membership is lower and has declined at a fasterrate from 1983 to 2017 relative to non-RTW states. Nonetheless, I detail belowseveral empirical steps to ensure that this study can appropriately assess theeffects of union membership on the earnings returns to RTI, independent of therelationship between union membership and technological change.

Third, and relatedly, this study does not measure whether union membershipattracts or repels technological innovation. This is a potential limitation of thisstudy and a relationship that should continue to be investigated in future work.3

On one hand, unionized routine workers could attract greater technologicaladoption (Acemoglu 2002). In an optimistic take on this scenario, the routineworkers use the technological innovations to enhance their productivity and, inturn, their earnings. That routine workers, rather than higher-pay and “higher-skill” workers, would primarily benefit from the onset of new technology,however, would be inconsistent with both the SBTC and RBTC literatures (Autorand Dorn 2013; Goldin and Katz 2008). More in line with prior evidence isthe more pessimistic scenario: unionized routine workers allow technologicalinnovation, and perhaps even attract it due to their higher relative labor costs, butnegotiate the implementation of the new technologies in a way that protects theirrelative earnings and employment (e.g., Fernandez 2001). Finally, an alternativeargument, based on the capital hold-up literature, suggests that firms are lesslikely to invest in new technology when labor is stronger, as the returns to theirinvestments would be shared among employees in the form of higher wages(Card, Devicienti, and Maida 2014; Cardullo, Conti, and Sulis 2015). Ultimately,this study does explicitly test which of these pathways is more plausible andtherefore refrains from ascertaining on which is more likely. Instead, this analysisnarrows in on the role of worker power in shaping the relative earnings trends ofroutine occupations, regardless of whether unions attract or repel technologicalinnovation in the first place. Results should be interpreted with this limitationin mind.

Data and Methods

I test the earnings and employment effect hypotheses on a sample of 3.3 millionemployed adults (ages 18–65) in the United States using the CPS MORG from1983 to 2017. The CPS MORG provides detailed data on the weekly earningsof employed adults, as well as data on each worker’s union membership.

I measure occupational earnings as the log of weekly earnings of employedadults. Weekly earnings are adjusted for inflation using the CPI and set to 2014USD. Self-employed and part-time workers (those working fewer than 30 hours

Dow



gmail.com

on 14 May 2020


per week) are excluded from the analysis. I exclude workers reporting less than$1 per hour in earnings. I multiply the earnings of top-coded earners by 1.4 in theprimary analysis, following recommendations in Denice and Rosenfeld (2018)and Lemieux (2006). Results without the top-code adjustments produce similarfindings. Following standard practice when using union data in the MORGsamples, I exclude allocated earners from the data, as the CPS does not includeunion status in its allocation model (Brueckner and Neumark 2014; Denice andRosenfeld 2018). Following Rosenfeld (2014) and Denice and Rosenfeld (2018),I exclude data from 1994 and 1995 due to a lack of valid allocation flags duringthose years. I use recommendations from Hirsch and Schumacher (2004) toidentify allocated earners in other years. Descriptive statistics are presented inAppendix B (table B1).

In the primary earnings analyses, I measure union status first at the individuallevel and, second, at the state-industry level using 3-year rolling averages ofunion coverage within each industry in each state. In a sensitivity check, Ialso produce estimates using state-industry union data only among high RTIoccupations or those in the top one-third of the RTI distribution. Doing so helpsto sidestep a potential mechanical relationship between technological changeand union membership. In the first approach, levels of union coverage withina state-industry may decline as employment shares of machine operators declineif union coverage is particularly concentrated among machine operators (asone example). In the sensitivity check, I measure the coverage among machineoperators; thus, even as employment shares of operators decline, the share of theremaining workers who are covered need not experience a mechanical decline.The results of the sensitivity check, presented in the Supplementary Material(table S1), align closely with the results from the primary analysis.

I follow Autor and Dorn (2013) in measuring the RTI of all occupationswithin the dataset. The RTI index has become the standard practice for mea-suring the task content of occupations (Goos et al. 2014:2511). Routine tasksconsist of repeated sequence of actions and are more easily replaceable throughtechnological innovations. A higher RTI score thus indicates that an occupationis more “routine task-intensive (RTI)” and, thus, can more readily be automated.In Appendix A, I provide more information on the construction of the RTI index.

I use the time-consistent occupation codes from Autor and Dorn (2013)to merge in the RTI data. The crosswalks that Autor and Dorn apply aremeant to allow for an examination of near-identical occupations over time(despite changes to occupational coding schemes within the CPS), but theyremain imperfect. Similar to Mishel et al. (2013), I find that inconsistencies inthe descriptive data appear in 2003, the year in which the revised occupationcodes were introduced. From 1983 to 2002, the mean RTI of occupations fallsfrom 1.27 to 0.96. In 2003, however, the mean RTI of occupations increasesfrom 0.96 to 1.04 before continuing its steady decline. As discussed in theEstimation Strategy and Sensitivity Checks below, I take several steps to accountfor the inconsistencies in the data due to the 2003 change in occupationcodes.

Dow



gmail.com

on 14 May 2020

https://academic.oup.com/sf/article-lookup/doi/10.1093/sf/soaa032#supplementary-data

10 Social Forces

Estimation Strategy

I first test whether higher union coverage within a state-industry in the UnitedStates does, indeed, contribute to higher earnings returns to an occupation’sRTI. This would be consistent with the earnings effect hypothesis. To test this,I begin with a simple estimation of trends in the returns to an occupation’s RTIwhile controlling for individual characteristics that likely affect labor supply.This estimation is modeled as follows:

log (Earn)jnst = β1RTIj + β2(RTIj • Yeart

) + β3Xj + β4(RTIj • Post02t

)

+ αn • αs • αt + εjnst. (1)

The outcome variable is the log weekly earnings of an individual (j) working inindustry (n) in a state (s) and year (t). RTI represents the RTI of the occupation,scaled from 0 (least routine) to 1 (most routine). The Yeart variable includedin the interaction term (see β2) represents a linear time trend, following similarspecifications from and Goos et al. (2014) and Mahutga, Curran, and Roberts(2018). Thus, if β2 is positive and statistically significant, this would implythat the earnings returns to higher RTI are increasing over time, independentof composition effects, such as a rise in educational attainment. Vector Xcontrols for demographic features (age, the square of age, education level, sex,marital status, a binary indicator of whether the worker lives in the city center,and race/ethnicity in the form of binary indicators for Black, Hispanic/Latino,or other non-White). The inclusion of these controls is standard in earningsestimates and is meant to account for other factors beyond unionization thatmight affect individual earnings. To account for the bias in levels of RTIintroduced during the 2003 occupation code changes, I include an interactionfor RTI and a dummy variable (Post02t) indicating whether the given year comesbefore or after the 2003 coding change. In the Supplementary Material (table S2),I also replicate all analyses segmenting the data into two sets of years: 1983–2002 and 2003–2017. The results are consistent with the primary analyses.Finally, the model also includes state-industry-year fixed effects. In practice,this takes into account that earnings trends are likely to vary across state andindustry. The main effect of any state-level variables (such as union coverage,but including other factors, such as minimum wages changes) is accountedfor in the state-year effects. I apply robust standard errors clustered at thestate level.4

I then extend equation (1) to measure the moderating role of union coverageon the trends in earnings returns to an RTI’s occupation. I first operationalizeunion coverage at the individual level using worker-level data on union member-ship. Given evidence that higher union membership also boosts the earnings ofnon-union members (Denice and Rosenfeld 2018), I then operationalize unioncoverage as the mean of state-industry-year union membership, as discussedearlier.

Dow



gmail.com

on 14 May 2020



Adding in union coverage creates a three-way interaction between RTI, thelinear year trend, and union coverage:

log (Earn)jnst = β1RTIj + β2(RTIj • Yeart

) + β3(RTIj • Yeart • Unionst

)

+ · · · + β4Xj + β5(RTIj • Post02t

) + αn • αs • αt + εjnst. (2)

Note that the three-way interaction in equation (2) also creates interactionsbetween union and year, as well as RTI and union; I conceal these from theformal estimation above solely for brevity. The primary coefficient of interestbecomes β3. A positive and significant slope would suggest that union coverageis associated with more favorable earnings returns to an occupation’s RTI overtime. Put differently, such a finding would support the hypothesis that higherunion coverage propels the relative earnings growth of jobs at greater risk ofautomation. This finding would corroborate the study’s primary hypothesis.

I then turn toward testing for a possible employment effect. To test whetherhigher union coverage leads to accelerated declines in occupations at greater riskof automation, I estimate the following:

RTIjnst = β1 (Unionst • Yeart)+ β2Xj +β3Post02 +αn •αs +αs •αt + εjnst. (3)

RTIjnst is again an occupation’s RTI. Yeart represents the linear time trend.Vector X controls for the same demographic features as before. Union mem-bership is measured again at the state-industry-year level. If β1 is negative andsignificant, this would suggest that rising union coverage within a state-industryhastens the decline in the employment shares of occupations at greater riskof automation, independent of compositional changes. Put differently, such afinding would suggest that more resilient union membership contributes to anaccelerated decline in the share of occupations with higher RTI.

Findings

I first present descriptive findings on trends in the returns to RTI across state-industries with below-average and above-average union coverage. Though myhypotheses and estimation strategy focus on the effect of within-state and within-industry variation in levels of union coverage on the returns to an occupation’sRTI, it is instructive to first examine this relationship across states to see if theanticipated patterns exist.

Figure 1 presents the earnings returns to RTI from 1983 to 2017 dependingon the state-industry’s union membership. State-industries with above-averageunion membership in the given year are presented in the black line, while thosewith below-average union membership are presented in the lighter-colored line.Two main patterns stand out from the figure. First, the earnings returns toan occupation’s RTI are consistently lower in states with below-average unionmembership. Put differently, an occupation at greater risk of automation is likelyto receive lower relative earnings if working in a state with lower levels of

Dow



gmail.com

on 14 May 2020

12 Social Forces

Figure 1. Earnings returns to RTI by union membership of state-industry

Note: Figure displays the marginal effects of an occupation’s RTI on log weekly earnings.

unionization. In 2017, the high- and low-union states differed in returns toRTI by a factor of two (an average return of −0.21 versus −0.45 log points,respectively, for each percentage point increase in an occupation’s RTI).

Second, the trends differ between the two groups. In state-industries withabove-average union membership, the earnings returns to RTI decline from−0.15 to −0.22 log points (−0.07 log point change) from 1983 to 2017. In state-industries with below-average union membership, the decline is much steeper.From 1983 to 2017, the earnings returns to RTI fall from −0.23 to −0.46log points, a −0.23 log point change. Whether examining levels or trends, thedescriptive evidence suggests that occupations at greater risk of automationexperience higher relative earnings in a state-industry with above-average levelsof unionization.

Do these relationships tend to hold when we look within each state-industryover time? To further investigate the role of unionization on the earningsreturns to an occupation’s RTI, I now turn toward the formal estimates of thehypotheses.

Table 1 presents the results of equations (1) and (2). Model 1 looks at whetherthe earnings returns to an occupation’s RTI have declined, on average, from1983 to 2017. The interaction term of RTI and the linear time trend (“year”) isnegative and statistically significant, indicating that the returns to RTI have, on

Dow



gmail.com

on 14 May 2020


Table 1. Estimations of Earnings Trends of Occupations by RTI and Level of Union Coverage(1983–2017)

(1) Trends (2) Individualunion

(3) State-industryunion

(4) State-industryunion on non-union

RTI −.217∗∗∗ (.001) −.221∗∗∗ (.010) −.249∗∗∗ (.016) −.249∗∗∗ (.017)

Union .115∗∗∗ (.001)

RTI # year (linear) −.006∗∗∗ (.001) −.006∗∗∗ (.001) −.004∗∗∗ (.001) −.002∗∗ (.001)

RTI # union .098∗∗∗ (.021) .048∗∗∗ (.012) .056∗∗ (.018)

Union # year (linear) −.003∗∗∗ (.000)

RTI # union # year(linear)

.003∗∗∗ (.001) .005∗∗∗ (.000) .006∗∗∗ (.000)

Observations 3,246,862 3,246,862 3,246,862 2,649,962

All models include individual-level controls (age, age squared, race/ethnicity, education, sex,marital status, and urban/non-urban dummy), state-industry-year fixed effects, and an interac-tion of RTI and a pre-/post-2003 dummy. Main effects of linear time trend and state-industry unionmembership omitted due to the use of state-year-industry fixed effects. Robust standard errorsin parentheses. Standardized coefficients presented for union coverage. ∗p < .05, ∗∗p < .01,∗∗∗p < .001.

average, fallen with each year. Put differently, high RTI occupations tend to seeless favorable earnings growth relative to lower RTI occupations, all else equal.

To what extent can within-industry variation in union coverage help toexplain differences in the earnings returns to an occupation’s RTI over time?Model 2 presents the results of the three-way interactions using individual-levelunion membership. The positive and significant slope for the RTI, year, andunion coverage interaction suggests that union membership is associated withgreater relative earnings growth for occupations with higher RTI. In other words,declining union membership is associated with declining earnings returns to anoccupation’s RTI.

Model 3 presents the results while operationalizing union membership atthe state-industry level for each year. Given evidence of union spillover effectsto non-union workers, as hypothesized, higher union coverage within a state-industry is associated with even greater relative earnings growth for occupationswith higher RTI. Declining levels of union coverage within a state-industry areassociated with declining earnings returns to an occupation’s RTI. Specifically, a1 SD increase in levels of union coverage is associated with a 0.5 percent increase,on average, in the earnings returns to an occupation’s RTI.5

Model 4 shows why measuring state-industry union membership, rather thanindividual union membership, is particularly useful. The results here repeat thesame estimate as Model 3, but only including non-union workers. Consistentwith results from Denice and Rosenfeld (2018), the findings suggest that higherunion membership at the state-industry level is beneficial for the earnings returnsto an occupation’s RTI even among the non-unionized.

To visually depict the magnitude of the results in table 1, Figure 2 shows themarginal effects of RTI on weekly earnings by year and union coverage. Themarginal effects are estimated using the results from Model 3 in table 1. Each

Dow



gmail.com

on 14 May 2020

14 Social Forces

Figure 2. Marginal effects of RTI on log weekly earnings by year and state-industry unionmembership

Note: Marginal effects from Model 3 of Table 1

year is labeled on the X-axis, while the Y-axis depicts the marginal effect of RTIon log weekly earnings. The three lines show how the marginal effects vary bylevel of union membership within the state-industry. The top line shows thatthe marginal effect of RTI on earnings actually increases over time when unionmembership is two SDs above the mean. In 2017, in fact, occupations at greaterrisk of automation are estimated to face no earnings penalty relative to lowerRTI occupations, on average, if union coverage was at two SD above the meanor higher.

In contrast, occupations at risk of automation fare much worse when state-industry union membership is low. At two SD below the mean, the penaltyassociated with an occupation’s RTI worsens over time. Moreover, the differencein returns to RTI for occupations in high-union versus low-union state-industrieswidens with each year. If a state-industry’s union density falls from two SDabove the mean, we can expect a 0.5 log point decline, on average, in earningsreturns to an occupation’s RTI in 2017. Falling from the mean to two SD belowthe mean results in additional 0.5 log point decline. These results support thisstudy’s primary hypothesis that higher union coverage within a state-industry isbeneficial for the relative earnings trends of higher RTI occupations.

To further contextualize the results, figure 3 presents a counterfactual evo-lution of the earnings returns to an occupation’s RTI if the United States had

Dow



gmail.com

on 14 May 2020


Figure 3. Counterfactual returns to an occupation’s RTI had union coverage in the UnitedStates remained constant from 1983 to 2017

Note: Marginal effects of RTI on log earnings. Findings from Table 1, Model 3.

experienced no decline in union coverage among high RTI occupations from1983 to 2017. In this scenario, union coverage would have remained steady atan average of 26.2 percent rather than dropping to an average of 13.7 percentby 2017. These counterfactual estimates are produced using the findings fromModel 3 in table 1, but with year dummies rather than linear time trends inthe interaction to account for any non-linear, year-to-year differences in relativeearnings growth rates. The solid black line in figure 3 represents the observedtrend in the marginal effects of RTI on earnings. The observed trend shows adecline in returns to RTI from 1983 to 2017. Consistent with prior evidenceof occupational earnings polarization, the observed trend shows a steep declinefrom 1990 to 2000, but no statistically significant difference from 2000 to 2017.

The dashed line represents the counterfactual earnings trends if levels of unioncoverage had not declined from 1983 levels. As opposed to the 14 percentagepoint decline in the observed trend from 1983 to 2017, the counterfactualtrend suggests that the earnings returns to an occupation’s RTI might not havefallen between 1983 and 2017 had union coverage remained stable from 1983onward. Moreover, the period of occupational earnings polarization in the 1990smight not have occurred in the counterfactual scenario. These counterfactualsare partial-equilibrium estimates and do not take into account other changesthat might occur if union coverage remained constant. The findings should be

Dow



gmail.com

on 14 May 2020

16 Social Forces

interpreted accordingly. Nonetheless, they suggest that the earnings trends ofoccupations at greater risk of automation are conditional on the strength oforganized labor.

In the Supplementary Material (figure S1), I present a version of figure 3that visualizes the returns to an occupation’s RTI over time conditional onindividual-level union membership, rather than state-level union membership,using the results from Model 2 of table 1. Compared to the 13.4 percentagepoint difference in the counterfactual and observed earnings trends when usingstate-industry union membership, the difference in the estimated earnings returnsto RTI for union versus non-union members is 9 percentage points. In thisalternative specification, the earnings returns to RTI for union members declineby 4.6 percentage points between 1983 and 2017, compared to 0.2 percentagepoints in the state-industry counterfactual. This trend is still more favorable thanthe 13.3 percentage point decline that non-union members experience. Thus, thisstudy’s primary hypothesis is supported whether investigating individual or state-industry union membership, but the effects are slightly stronger when analyzingstate-industry union membership.

For further evidence of the relationship with earnings polarization, figure 4plots the potential earnings gains by occupation type in 2017 had union coverageremained steady from 1983 onward (using the same estimates from table 1). Thefigure shows the estimated increase in earnings for the six primary occupationcategories featured in Autor and Dorn (2013). The occupations are lined fromthe lowest average pay in 1983 (beginning with service occupations on theleft side) to the highest average pay (management/professionals on the right).The findings show that the greatest increase in earnings would occur amongthe operator occupation groups, which includes machine operators and similarhigh RTI occupations. In the absence of union decline from 1983 to 2017, theestimates suggest that the operator groups would see an increase in earningsof about 1.5 log points, higher than the other occupation groups. Notably, theoccupations benefiting the most are in the middle of the earnings distribution.Thus, an increase in their earnings acts as the inverse of earnings polarization.Again, the evidence suggests that assessments of the earnings growth of high RTIoccupations, as well as theories regarding the sources of earnings polarization inthe 1990s, must take into account the declining strength of organized labor. Hadunion coverage remained constant from 1983 onward, rather than declining byaround 50 percent, occupations at greater risk of automation would likely faremuch better in terms of their relative earnings growth.

Evidence of an Employment Effect?

The evidence presented thus far supports the hypothesis that stronger unionsinhibit a decline in the earnings returns to an occupation’s RTI. But, do thegains in relative earnings come at a cost? Specifically, do we find that strongerunion coverage within a state-industry is associated with an accelerated declinein the share of occupations at risk of automation? Table 2 presents the results ofequation (3), which tests for a possible employment share effect.

Dow



gmail.com

on 14 May 2020



Figure 4. Estimated increase in earnings by occupation type in 2017 had union coverage inthe United States not declined from 1983 levels

Note: Occupation groups sorted from lowest (left) to highest (right) mean in wage in 1983.Y -axis: gain in log earnings if union membership remained constant from 1983 to 2017.

Table 2. OLS Estimates of Effect of State-Industry Union Coverage on Trends in RTI of EmployedAdults

Outcome: RTI of employedadults

(1) (2)

Year (linear) −.004∗∗∗ (.000) −.010∗∗∗ (.001)

Union coverage −.064∗∗∗ (.008)

Union coverage # year (linear) −.001∗∗ (.000)

Observations 3,285,672 3,285,672

Dependent variable: RTI. All models include individual-level controls (age, age squared,race/ethnicity, education, sex, marriage status, and urban/non-urban dummy), state-industryfixed effects, state-year fixed effects, and an interaction of RTI and a pre-/post-2003 dummy.Robust standard errors in parentheses. Standardized coefficients presented for collectivebargaining coverage. ∗p < .05, ∗∗p < .01, ∗∗∗p < .001.

Model 1 first looks at trends in the RTI of all occupations within state-industries. As expected, the linear time trend is negative and significant. Thisindicates that occupations at greater risk of automation are, indeed, losingemployment shares over time relative to less routine occupations. Model 2 addsthe interaction of union coverage and the linear time trend. If the interaction were

Dow



gmail.com

on 14 May 2020

18 Social Forces

Figure 5. Counterfactual change in mean RTI of occupations had union coverage not declinedfrom 1983 to 2017

to be negative and significant, this would suggest that the higher union coveragewithin a state-industry contributes to faster decline in employment shares of highRTI occupations within a state-industry. The evidence supports this conclusion.

To contextualize this finding, figure 5 displays the same counterfactual asapplied in the earnings estimates. Had union coverage remained constant from1983 to 2017, to what extent would the mean RTI of occupations have declined?The observed trend (solid line) shows that the mean RTI of occupations fell byaround 28 percent from 1983 to 2017. Had union coverage remained constant, itwould have fallen by around 34 percent instead, according to the estimates. Thedifference is 6 percentage points, or around 21 percent steeper than the observeddecline. The results support the existence of a negative employment share effect.

To summarize the results, when union coverage within a state-industry ishigher, occupations at greater risk of automation experience greater earningsgrowth, but also tend to experience accelerated declines in employment shares.Returning to the language of the RBTC literature, we can thus conclude thatorganized labor inhibits the polarization of occupational earnings, but alsoappears to accelerate the polarization of occupational employment shares.

Sensitivity Checks

I present several sensitivity checks in the Supplementary Material to assessthe consistency of the study’s findings under alternative model specifications.

Dow



gmail.com

on 14 May 2020


First, I test whether the results hold when a state-industry’s union membershipamong high RTI occupations is substituted for state-industry union membershipamong all occupations. Following Lordan and Neumark (2017), high RTIoccupations are those defined in the top one-third of the RTI distribution. Theadvantage of such an approach is the sidestepping of a potential mechanicalrelationship between technological change and the decline of union membership.If union membership is concentrated among routine occupations, and automa-tion affects routine occupations more than occupations, then automation maydrive down levels of union membership. When measuring union membershipamong high RTI occupations, I am capturing the mean unionization amongthe high RTI occupations that remain, regardless of their overall employmentshares. The results, as shown in table S1, are consistent with the study’s primaryfindings.

As an additional check, I re-estimate the primary models (operationalizingthe mean union density of all occupations in the state-industry each year)while controlling for the effect of trends in the mean RTI of occupationsin a state-industry on earnings. This helps to ensure that the effect ofunion density on returns to an occupation’s RTI over time can be inter-preted independent of changes in the composition of occupations within theindustry. As shown in table S3, the results are consistent with the primaryanalysis.

In the primary analysis, I include a pre-/post-2003 dummy interacted with anoccupation’s RTI to ensure that the change in occupational coding in 2003 doesnot affect this study’s results. As a sensitivity check, I also re-estimate the resultson both the pre- and post-2003 samples (the 1983–2002 samples and the 2003–2017 samples). The results, presented in table S2 in the Supplementary Material,show that the primary findings hold in both the segmented samples.

Finally, this study measures the effect of variation in union membershipwithin state-industries on earnings trends, but we could also look at the effectof variation in union membership between state-industries in a single year. Across-sectional, multilevel model estimated in a single year is more capable oftaking advantage of between-state variation in union coverage, but at the costsof increased likelihood of omitted variable bias and the loss of all-time effects.Nonetheless, I present cross-sectional results of the effect of union coverageon earnings trends in table S4. In line with this study’s primary results, thecross-sectional models suggest that higher union coverage in a state-industrycontributes to greater earnings returns to an occupation’s RTI.

Discussion

RBTC has emerged as the dominant explanation for the decline in relativeearnings of occupations at greater risk of automation. In contrast, this studydraws on power resource theory and the comparative institution literatureto posit that the differential earnings trends of high-RTI occupations can belargely attributed to the decline of organized labor. Placing the RBTC framework

Dow



gmail.com

on 14 May 2020





20 Social Forces

into political-institutional context follows past research that has demonstratedhow differences in labor market institutions directly affect the distribution ofresources (Fernandez 2001; Kristal 2013). Rather than focusing on broaderpatterns of earnings inequality, however, this study narrows in on the earningsand employment penalties associated with an occupation’s RTI—two empiricalbenchmarks that are central to the emergent literature on automation, routinework, and occupational polarization.

Using micro-data spanning 3.3 million employed adults from 1983 to 2017,this study finds that higher levels of union coverage within a state and industry inthe United States contribute to higher earnings growth for routine occupations.Thus, the observed decline of union coverage in the United States appearsto have contributed to the period of occupational wage polarization in the1990s, as well as the general decline in the earnings returns to an occupation’sRTI. However, the relative earnings gains perhaps come at the cost of anaccelerated decline in employment shares for routine occupations. I discuss thesetwo countervailing effects—an earnings effect and employment share effect—inturn.

In prior literature related to RBTC, the decline in relative earnings growthamong occupations at greater risk of automation is primarily attributed to exoge-nous forces shifting demand away from routine tasks. These past studies havebeen influential in shaping perceptions around the economic fate of middle-skill,middle-pay workers, yet tend to neglect contextual factors that might impedea straightforward relationship between technological change and variation inoccupational earnings trends. Consistent with power resource theory and priorresearch on the union earnings premium, this study finds that the strengthof organized labor shapes changes in the earnings distribution. More so, thefindings show that organized labor is particularly consequential for the earningstrends of occupations at greater risk of automation. When unions are strongerwithin a state and industry, high RTI occupations experience greater relativeearnings growth, all else equal. This is true even for non-unionized workers,consistent with results from Denice and Rosenfeld (2018). The counterfactualsimulations show the magnitude of the results: had unionization remainedstable from 1983 (an average of 26.2 percent), rather than dropping to anaverage of 13.7 percent by 2017, the earnings returns to an occupation’s RTImight not have declined between 1983 and 2017, rather than experiencinga steep decrease. Moreover, the period of occupational earnings polarizationin the 1990s might not have occurred. The counterfactual results are robust,albeit with slightly weaker effects, when operationalizing union membershipat the individual rather than state-industry level (see figure S1). These findingssuggest that the RBTC framework, which attributes the declining relativeearnings of routine occupations primarily to technological advancements, over-looks the importance of worker power in shaping the earnings trends of suchoccupations.

Despite the positive earnings effect, the evidence also points to a negativeemployment share effect of higher union membership. Consistent with priorevidence of organized labor inducing earnings-employment trade-offs, this study

Dow



gmail.com

on 14 May 2020



finds that higher union coverage within a state-industry can achieve higherrelative earnings growth for automatable jobs, but at the cost of an accelerateddecline in employment shares of automatable jobs. Had union coverage remainedconstant from 1983 to 2017, the mean RTI of occupations would have fallen byaround 34 percent instead of 28 percent, a 6 percentage point (or 21 percent)steeper decrease than the observed decline.

These findings have direct implications on the emerging literature aroundautomation and its consequences for labor markets and the income distribution.Most directly, they question the extent to which automation deserves credit forthe declining relative earnings of routine occupations. Declining demand forroutine tasks cannot, on its own, explain the observed decline in the earningsreturns to an occupation’s RTI in the United States from 1983 to 2000, orthe lackluster earnings growth in among high RTI occupations from 2000onward. Instead, this study makes clear that the decline of organized labor isa central source of the declining relative earnings growth of jobs at greaterrisk of automation. At the same time, organized labor may act to accelerateemployment polarization, or the relative decline of high RTI occupation shares.This may help to explain why, as mentioned in this study’s Introduction, manycountries in the European Union (which tend to feature higher and more resilientlevels of collective bargaining coverage than the United States) have experiencedemployment polarization without earnings polarization, though future researchshould investigate this in more detail.

Several limitations and opportunities for future research should be acknowl-edged. First, this study uses repeated cross sections from the CPS MORG tounderstand how worker power shapes the relative earnings trends of routineoccupations. Future work, however, would greatly benefit from the use ofpanel data, such as that available in the Panel Study of Income Dynamics, toassess whether the employment and earnings trajectories of routine occupationsare conditional on union membership. This would allow for a more detailedinvestigation of the potential employment effects of union membership amongroutine occupations. Second, this study does not engage with the potential het-erogeneous effects of union membership across different demographic profiles.Prior research shows, however, that unions have been particularly beneficial forthe earnings of non-White workers (Rosenfeld 2014). It may thus be the casethat the observed decline in relative earnings of routine occupations, spurredin part by declining worker power, contributes to greater racial inequality inearnings and poverty outcomes. Future work should investigate the implicationsof technological change on racial differences in economic opportunity in moredetail. Third, this study does not assess whether unions, or other labor marketinstitutions, attract or dispel greater investment into new technology. Futurework should continue to improve our ability to measure technological changeand to more explicitly measure whether greater worker power is a magnet orrepellant for greater technological innovation.

In general, researchers should continue to place analyses of automation andoccupational change into political-institutional context. As this study demon-strates, the social consequences of technological change appear to be conditionalon the strength of organized labor.

Dow



gmail.com

on 14 May 2020

22 Social Forces

Appendix ACalculating the RTI Index

Using information on job tasks from the Occupational Information Network(O∗NET) database, the RTI index assigns values to occupations according tothe extent to which they require routine, manual, or abstract tasks. Specifically,Acemoglu and Autor (2011) compute an occupation’s RTI as follows:

RTIk = ln(TR

k

)− ln

(TM

k

)− ln

(TA

k

), (A.1)

where TRk , TM

k , and TAk represent the level of routine, manual, and abstract

task inputs, respectively, for an occupation k. Routine tasks consist of repeatedsequence of actions and are more easily replaceable through technologicalinnovations. Manual tasks, meanwhile, do not generally follow a predictablesequence and therefore are more resistant to automation. Abstract tasks tend tocomplement new technologies rather than to be automated. Given this equation,a higher value of RTIk indicates that an occupation is more “RTI” and, thus, canmore readily be automated.

Appendix BDescriptive Statistics

Table B1. Descriptive Statistics

Variable Mean St. Dev Min. Max.

State-industry-year union membership 0.17 0.14 0.00 0.68

Female 0.46 0.50 0 1

Age 39.2 11.6 18 65

Education: low (HS degree) 0.44 0.50 0 1

Education: high (college degree) 0.30 0.46 0 1

Race/Eth.: White 0.86 0.35 0 1

Race/Eth.: Black 0.09 0.28 0 1

Race/Eth.: Other 0.04 0.19 0 1

Race/Eth.: Hisp 0.09 0.29 0 1

Married 0.61 0.48 0 1

City Center 0.23 0.42 0 1

RTI (scaled zero to one) 0.33 0.19 0 1

Notes1. Advanced technologies tend to benefit owners of capital and high-skill

workers which, combined with increased consumption of goods and services

Dow



gmail.com

on 14 May 2020


among higher-paid workers, is theorized to contribute to a rise of employ-ment shares among low-skill occupations, typically in the services sector.Some scholars, however, disagree on the extent to which advancements intechnology deserve credit for employment polarization (Salvatori, 2015).

2. Company data show that Hyundai operates a plant in Alabama; Nissanoperates plants in Mississippi and Tennessee; and Toyota operates plantsin Kentucky, Texas, Mississippi, Alabama, and Indiana.

3. One challenge in measuring the association between union membership andinvestments into new technology is the lack of a comprehensive measure oftechnological investment at the state and industry level. The routine taskintensity approach, as adopted in this paper, is recognized as “the best wayto capture the impact of recent technological progress” (Goos, Manning, andSalomons 2014: 2511). However, detailed state-industry level data on firminvestments into ICT would be beneficial in investigating this relationshipmoving forward.

4. Results are consistent when standard errors are clustered on both occupationand state.

5. The inclusion of state-year-industry fixed effects accounts for other time-varying institutional factors at the state-industry level that affect the rela-tionship between union membership and earnings of routine occupations.Nonetheless, this study cannot rule out the possibility that changes in state-industry union membership are closely correlated with changes in otherstate-industry characteristics.

Supplementary Material

Supplementary material is available at Social Forces online.

About the AuthorsZachary Parolin is a postdoctoral research scientist with Columbia University’sCenter on Poverty and Social Policy. His research focuses on the role of policy,politics, and worker power in shaping patterns of social inequality across high-income countries. Recent publications at Demography, Socioeconomic Review,and elsewhere focus on the effects of social and labor market policies on levelsof poverty and income equality in the United States and European Union.

ReferencesAcemoglu, Daron. 2002. "Directed Technical Change." The Review of Economic Studies 69(4):781–809. doi:

10.1111/1467-937x.00226.Acemoglu, Daron and David Autor 2011. “Skills, Tasks and Technologies: Implications for Employment &

Earnings.” Handbook of Labor Economics, Volume 4b .Ahlquist, John S. and Margaret Levi 2013. In the Interest of Others: Organizations and Social Activism.

Princeton: Princeton University Press.

Dow



gmail.com

on 14 May 2020


https://doi.org/10.1111/1467-937x.00226

24 Social Forces

Antonczyk, Dirk, Thomas DeLeire and Bernd Fitzenberger 2010. “Polarization and Rising Wage Inequality:Comparing the U.S. and Germany.” IZA Discussion Paper Series No. 4842.

Autor, David, Frank Levy and Richard J. Murnane. 2003. "The Skill Content of Recent Technological Change:An Empirical Exploration." Quarterly Journal of Economics, 118 (4) (November):1279–334.

Autor, David and David Dorn. 2013. "The Growth of Low-Skill Service Jobs and the Polarization of the UsLabor Market." American Economic Review 2013(5):1553–97.

Biegert, Thomas. 2017. "Welfare Benefits and Unemployment in Affluent Democracies: The ModeratingRole of the Institutional Insider/Outsider Divide." American Sociological Review 82(5):1037–64. doi:10.1177/0003122417727095.

Böhm, Michael J. 2017 The Price of Polarization: Estimating Task Prices under Routine-Biased TechnicalChangeIZA (Institute for the Study of Labor) Discussion Paper Series11220

Brady, David 2009. Rich Democracies, Poor People: How Politics Explain Poverty . Oxford: Oxford UniversityPress.

Brady, David, Regina S. Baker and Ryan Finnigan. 2013. "When Unionization Disappears: State-LevelUnionization and Working Poverty in the United States." American Sociological Review 78(5):872–96.doi: 10.1177/0003122413501859.

Brady, David, Agnes Blome and Hannah Kleider 2016. “How Politics and Institutions Shape Poverty andInequality”. In The Oxford Handbook of the Social Science of Poverty , edited by Brady, David, Burton,Linda M. Oxford: Oxford University Press.

Brenke, Karl, Ulf Rinne and Klaus F. Zimmermann 2011. “Short-Time Work: The German Answer to theGreat Recession.” IZA (Institute for the Study of Labor) Discussion Paper Series No. 5780.

Brueckner, Jan K. and David Neumark. 2014. "Beaches, Sunshine, and Public Sector Pay: Theory andEvidence on Amenities and Rent Extraction by Government Workers." American Economic Journal:Economic Policy 6(2):198–230. doi: 10.1257/pol.6.2.198.

Card, David, Francesco Devicienti and Agata Maida. 2014. "Rent-Sharing, Holdup, and Wages: Evidencefrom Matched Panel Data." The Review of Economic Studies 81(1):84–111. doi: 10.1093/restud/rdt030.

Cardullo, Gabriele, Maurizio Conti and Giovanni Sulis 2015. “Sunk Capital, Unions and the Hold-Up Problem:Theory and Evidence from Cross-Country Sectoral Data.” European Economic Review 76:253–74. doi:10.1016/j.euroecorev.2015.03.005.

Cortes, Guido Matias. 2016. "Where Have the Middle-Wage Workers Gone? A Study of Polarization UsingPanel Data." Journal of Labor Economics 34(1):63–105.

Denice, Patrick and Jake Rosenfeld 2018. “Unions and Nonuion Pay in the United States, 1977-2015.”Sociological Science 5:541–61.

DiNardo, John Enrico, Nicole M. Fortin and Thomas Lemieux. 1996. "Labor Market Institutions and theDistribution of Wages, 1973-1992: A Semiparametric Approach." Econometrica 64(5):1001–44.

Dwyer, Rachel E. 2013. "The Care Economy? Gender, Economic Restructuring, and Job Polarization in theU.S. Labor Market." American Sociological Review 78(3):390–416. doi: 10.1177/0003122413487197.

Fernandez, Roberto M. 2001. "Skill-Biased Technological Change and Wage Inequality: Evidence from aPlant Retooling." American Journal of Sociology 107(2):273–320. doi: 10.1086/324009.

Firpo, Sergio Nicole M. Fortin Thomas Lemieux 2011 Occupational Tasks and Changes in the WageStructureIZA (Institute for the Study of Labor) Discussion Paper Series5542

Goldin, Claudia Dale and Lawrence F. Katz 2008. The Race between Education and Technology . Cambridge,MA; London: Belknap.

Goos, Maarten, Alan Manning and Anna Salomons. 2009. "Job Polarization in Europe." American EconomicReview: Papers & Proceedings 99 2:58–63.

Dow



gmail.com

on 14 May 2020

https://doi.org/10.1177/0003122417727095

https://doi.org/10.1177/0003122413501859

https://doi.org/10.1257/pol.6.2.198

https://doi.org/10.1093/restud/rdt030

https://doi.org/10.1016/j.euroecorev.2015.03.005

https://doi.org/10.1177/0003122413487197

https://doi.org/10.1086/324009


Goos, Maarten, Alan Manning and Anna Salomons. 2014. "Explaining Job Polarization: Routine-BiasedTechnological Change and Offshoring." American Economic Review 104(8):2509–26.

Hertel-Fernandez, Alexander 2019. State Capture: How Conservative Activists, Big Businesses, andWealthy Donors Reshaped the American States – and the Nation. New York, NY: Oxford UniversityPress.

Hirsch, Barry T. and David A. MacPherson. 2003. "Union Membership and Coverage Database fromthe Current Population Survey: Note." Industrial & Labor Relations Review 56 (2):349–54. doi:10.2307/3590942.

Hirsch, B.T. 2004. “What Do Unions Do for Economic Performance?” Journal of Labor Research 25:417–55.Hirsch, Barry and Edward J. Schumacher. 2004. "Match Bias in Wage Gap Estimates Due to Earnings

Imputation." Journal of Labor Economics 22(3):689–722.Jacobs, David and Jonathan C. Dirlam. 2016. "Politics and Economic Stratification: Power Resources

and Income Inequality in the United States." American Journal of Sociology 122(2):469–500. doi:10.1086/687744.

Kalleberg, Arne L., Michael Wallace and Lawrence E. Raffalovich. 1984. "Accounting for LaborsShare: Class and Income Distribution in the Printing Industry." Ilr Review 37(3):386–402. doi:10.1177/001979398403700305.

Katz, Lawrence F. and David H. Autor 1999. “Changes in the Wage Structure and Earnings Inequality”. InHandbook of Labor Economics, vol 3a, edited by Ashenfelter, Orley, Card, David. Amsterdam: Elsevier.

Katz, Lawrence F. and Kevin M. Murphy. 1992. "Changes in Relative Wages, 1963-1987 - Supply-and-Demand Factors." Quarterly Journal of Economics 107 (1):35–78. doi: 10.2307/2118323.

Kimeldorf, Howard. 1992. "World War Ii and the Deradicalization of American Labor: The Ilwu as a DeviantCase." Labor History 33(2):248–78. doi: 10.1080/00236569200890121.

Korpi, Walter 1983. The Democratic Class Struggle. London: Routledge & Kegan Paul.Korpi, Walter. 1985. "Action, Resources and Power - Causal and Final Explanatory Models in the Analysis

of Power." Sociologisk Forskning 22(1):59–88.Kristal, Tali. 2013. "The Capitalist Machine: Computerization, Workers’ Power, and the Decline

in Labor’s Share within U.S. Industries." American Sociological Review 78(3):361–89. doi:10.1177/0003122413481351.

Lemieux, Thomas. 2006. "Increasing Residual Wage Inequality: Composition Effects, Noisy Data, or RisingDemand for Skill?" American Economic Review 96(3):461–98. doi: 10.1257/aer.96.3.461.

Leonard, Jonathon. 1992. "Unions and Employment Growth." Industrial Relations: A Journal of Economyand Society 31(1):80–94. doi: 10.1111/j.1468-232X.1992.tb00299.x.

Lindbeck, Assar and Dennis J. Snower. 2001. "Insiders Versus Outsiders." The Journal of EconomicPerspectives 15(1):165–88.

Lordan, Grace and David Neumark 2017. “People Versus Machines: The Impact of Minimum Wages onAutomatable Jobs.” National Bureau of Economic Research Working Paper Series No. 23667 . doi:10.3386/w23667.

Mahutga, Matthew C., Michaela Curran and Anthony Roberts. 2018. "Job Tasks and the ComparativeStructure of Income and Employment: Routine Task Intensity and Offshorability for the Lis." InternationalJournal of Comparative Sociology 59(2):81–109.

Mishel, Lawrence, John Schmitt and Heidi Shierholz 2013. “Assessing the Job Polarization Explanation ofGrowing Wage Inequality.” Economic Policy Institute Working Paper #295.

Naticchioni, Paolo, Giuseppe Ragusa and Riccardo Massari 2014. “Unconditional and Conditional WagePolarization in Europe.” IZA Discussion Paper Series No. 8456.

Dow



gmail.com

on 14 May 2020

https://doi.org/10.2307/3590942

https://doi.org/10.1086/687744

https://doi.org/10.1177/001979398403700305

https://doi.org/10.2307/2118323

https://doi.org/10.1080/00236569200890121

https://doi.org/10.1177/0003122413481351

https://doi.org/10.1257/aer.96.3.461

https://doi.org/10.1111/j.1468-232X.1992.tb00299.x

https://doi.org/10.3386/w23667

26 Social Forces

Newman, Robert J. 1983. "Industry Migration and Growth in the South." The Review of Economics andStatistics 65(1):76–86. doi: 10.2307/1924411.

Nickell, S.J. and M. Andrews 1983. “Unions, Real Wages and Employment in Britain 1951-79.” OxfordEconomic Papers 35:183–206.

OECD 2011. Divided We Stand: Why Inequality Keeps Rising. Paris: OECD.OECD 2017. OECD Employment Outlook 2017 . Paris: OECD Publishing.Pencavel, John H. 1984. "The Tradeoff between Wages and Employment in Trade Union Objectives∗." The

Quarterly Journal of Economics 99(2):215–31. doi: 10.2307/1885523.Rosenfeld, Jake 2014. What Unions No Longer Do. Cambridge, MA: Harvard University Press.Salvatori, Andrea. 2015. "The Anatomy of Job Polarization in the Uk." IZA Discussion Paper Series. Bonn:

Germany. No. 9193.Thelen, K. 2015. "Varieties of Liberalization and the New Politics of Social Solidarity (Vol 68, Pg 710, 2015)."

Ilr Review 68(4):970–70. doi: 10.1177/0019793915595009.VanHeuvelen, Tom. 2018. "Moral Economies or Hidden Talents? A Longitudinal Analysis of Union Decline

and Wage Inequality, 1973–2015." Social Forces 97(2):495–530. doi: 10.1093/sf/soy045.Visser, Jelle and Daniele Checchi 2011. “Inequality and the Labour Market: Unions.” The Oxford Handbook

of Economic Inequality .Western, Bruce and Jake Rosenfeld. 2011. "Unions, Norms, and the Rise in U.S. Wage Inequality." American

Sociological Review 76(4):513–37. doi: 10.1177/0003122411414817.Wilmers, Nathan. 2017. "Labor Unions as Activist Organizations: A Union Power Approach to Estimating

Union Wage Effects." Social Forces 95(4):1451–78. doi: 10.1093/sf/sow108.

Dow



gmail.com

on 14 May 2020

https://doi.org/10.2307/1924411

https://doi.org/10.2307/1885523

https://doi.org/10.1177/0019793915595009

https://doi.org/10.1093/sf/soy045

https://doi.org/10.1177/0003122411414817

https://doi.org/10.1093/sf/sow108

Documents

Automation,OccupationalEarningsTrends ...zachparolin.com/wp-content/uploads/2020/05/SocialForces_Parolin.pdfreviewers.I also appreciate the feedback of participants at the 2018 ESPANet